Methods to Estimate Respiratory Rate Using the

Photoplethysmography Signal

Ayalon Angelo de Moraes Filho

, Guilherme Schreiber

, Julio Alexander Sieg

Maicon Diogo Much

, Vanessa de Moura Bartoski

and César Marcon

School of Technology, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Brazil

Keywords: Health, Photoplethysmography, Respiratory Rate Estimation.

Abstract: Academia and industry have devoted significant effort to the research and development of smart wearable

devices applied to health monitoring. The photoplethysmography (PPG) sensor is widely used for monitoring

biosignals, such as heart and respiratory rate (RR), which are influenced by the cardiovascular system. This

work focuses on analyzing methods for RR estimation regarding the effect of breathing on the PPG signal

variation. This work describes, implements, and analyzes four methods for estimating RR. These methods are

based on capturing RR using Fast Fourier Transform, median, and extracting physiological characteristics

induced by respiration in the PPG signal. The most efficient method merges three RR calculations analyzed

on the same signal, achieving nearly 93% of efficacy in the best scenario. The method efficacies were

calculated using PPG signals from the BIDMC and CapnoBase databases collected from patients during

hospital care. The analysis allows for understanding and mitigating the RR estimation challenges and

evaluating the most efficacy method for a wearable device monitoring scenario.

1 INTRODUCTION

The aging population, the availability of mobile

broadband connectivity, and the development of

sophisticated technologies have driven the adoption

of personalized, digital, or remote patient monitoring

methods. This process was further accelerated with

the emergence of the coronavirus pandemic, which

increased pressure on limited hospital facilities,

requiring medical service providers to accelerate the

research and implementation of new technologies for

monitoring health outside the hospital (Olivadoti,

2022), especially in the patient's home.

Sensor innovations allow vital signs to be

measured with clinical-grade accuracy in a residential

setting. Wearable devices are more accessible,

enabling home monitoring of philological signs, such

as body temperature, heart rate, Respiratory Rate

(RR), blood pressure, and oxygen saturation.

RR is a valuable diagnostic and prognostic marker

of health. In hospital healthcare, it is a highly

sensitive marker of acute deterioration. For example,

https://orcid.org/0000-0001-7044-1504

https://orcid.org/0000-0002-1037-4853

https://orcid.org/0000-0003-3966-9855

an elevated RR predicts cardiac arrest and in-hospital

mortality and may indicate respiratory dysfunction.

Consequently, RR is measured between four and six

hours in hospitalized patients with acute illness. RR

is also used in emergency triage. In primary care, RF

is used to identify pneumonia and as a marker of

pulmonary embolism. However, RR is usually

measured by manually counting chest wall

movements (outside intensive care) and is a time-

consuming, imprecise, and poorly performed process

(Charlton et al., 2018).

The optical photoplethysmograph (PPG) sensor is

commonly found among wearable physiological

signal monitoring devices, such as pulse oximeters,

due to its simplicity, low cost, and non-invasive

approach. The PPG sensor is directly related to the

cardiovascular system, detecting blood content and

volume changes in the microvascular system.

Because the cardiovascular and respiratory systems

are correlated, researchers have made efforts to

develop algorithms capable of inferring the

respiration rate from the PPG signal.

https://orcid.org/0000-0002-1760-907X

https://orcid.org/0000-0001-6365-9273

https://orcid.org/0000-0002-7811-7896

Moraes Filho, A., Schreiber, G., Sieg, J., Much, M., Bartoski, V. and Marcon, C.

Methods to Estimate Respiratory Rate Using the Photoplethysmography Signal.

DOI: 10.5220/0011729100003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 445-452

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

445

The main objective of this work is to present RR

inference methods from PPG signals for obtaining the

best efficiency for absolute error equal to 0 breaths

per minute (rpm) in the analyzed data.

2 THEORETICAL FOUNDATION

PPG is used to measure blood volume changes in the

microvascular tissue bed under the skin; these

changes occur due to the pulsatile nature of the

circulatory system (Kamal et al., 1989). As an optical

technique, PPG requires a light source and a

photodetector. Light passing through biological tissue

can be absorbed by different substances, including

pigments in skin, bones, and arterial and venous

blood. Most changes in blood flow occur primarily in

the arteries and arterioles. PPG sensors optically

detect changes in blood flow volume (i.e., changes in

detected light intensity) in the microvascular tissue

bed, either through reflection or transmission through

the tissue (Tamura et al., 2014).

Figure 1 exemplifies a PPG waveform consisting

of direct current (DC) and alternating current (AC)

components. The DC component corresponds to the

transmitted or reflected optical signal detected in the

tissue; this component depends on the tissue structure

and the average volume of arterial and venous blood.

The AC component shows changes in blood volume

between the cardiac cycle's systolic and diastolic

phases; the AC component's fundamental frequency

depends on the heart rate and is superimposed on the

DC component (Tamura et al., 2014).

Figure 1: PPG waveform example (Tamura et al., 2014).

PPG pulse wave morphology is influenced by (i)

the heart, which considers cardiac ejection

characteristics, including heart rate and rhythm, and

stroke volume; (ii) circulation, including

cardiovascular properties such as arterial stiffness and

blood pressure; (iii) additional physiological

processes, including breathing and the autonomic

nervous system, that can be affected by stress; and

(iv) diseases (Mejıa-Mejıa et al., 2021). The quality

of the PPG signal depends on the wavelength of the

light, measurement location, i.e., sensor attachment

location, contact force, motion artifacts, the breathing

of the individual being measured, and ambient

temperature (Tamura & Maeda, 2018). These factors

generate various types of additive noise (artifacts)

that can be contained in PPG signals, affecting signal

characteristics.

Respiratory-induced changes in intrathoracic

pressure are transmitted to the central veins,

generating a change in blood pressure that the

spectrum of the PPG signal can detect. Breathing can

induce variations in the PPG signal in three ways

(Dehkordi, 2018):

 Respiratory-Induced Intensity Variation (RIIV) -

Changes in venous return due to changes in

intrathoracic pressure throughout the respiratory

cycle cause a modulation of the baseline (i.e., the

continuous component - DC) of the PPG signal;

 Respiratory-Induced Amplitude Variation

(RIAV) - During inspiration, the systolic volume

of the left ventricle decreases due to changes in

intrathoracic pressure, reducing the pulse

amplitude and the opposite happens during

expiration;

 Respiratory-Induced Heart Rate Variation

(RIFV) - Heart rate varies throughout the

respiratory cycle, increasing during inspiration

and decreasing during expiration.

3 RELATED WORK

Table 1 stands 19 works and the one proposed here

concerning (i) the year of publication, (ii) the

foundation that guides the extraction of the RR, (iii)

the extraction method used, (iv) the domain in which

the signals were analyzed; and (v) database used to

obtain the PPG signal.

The Base column of Table 1 shows that the RR

extraction from most works is based on the

physiological characteristics of breathing, with a

smaller portion of the works employing the PPG

signal filtering process. When we analyze the filtering

methods in isolation, represented by Fl, we realize

HEALTHINF 2023 - 16th International Conference on Health Informatics

446

Table 1: Related work comparison.

Article Base Method Domain Dataset

Pimentel et al., 2017 Rpc FFT*, RIIV, RIAV, RIFV, Fusion Freq. CapnoBase, BIDMC

hanidou, 2017 R

c EEMD, RIAV, RIFV Tem

o Ad hoc

Motin et al., 2018 Fl FFT, EEMD Fre

.Ca

noBase, MIMIC

Khreis et al., 2018 R

c RIIV, RIAV, RIFV Time Ca

noBase

Birrenkott et al., 2018 Rpc FFT*, RIIV, RIAV, RIFV Freq. CapnoBase, MIMIC

Motin et al., 2019 Fl EMD, EEMD, CEEMD, CEEMDAN, ICEEMDAN Time CapnoBase, MIMIC

Yang, 2019 Rpc FFT*, RIIV, RIAV, RIFV, Fusion Freq. Ad hoc

Pollreisz and Nejad,

2020

Rpc FFT*, RIIV, RIAV, RIFV, Fusion Freq. Ad hoc

Motin et al., 2020 Fl FFT*, EEMD, RIIV, RIAV, RIFV Freq. Ad hoc

Pollreisz and

TaheriNejad, 2020

Rpc FFT*, RIFV Freq. Ad hoc

Khreis et al., 2020 Rpc FFT*, RIIV, RIAV, RIFV Freq. CapnoBase, Sherpa

Lazazzera and Carrault,

2020

Fl, Rpc FFT*, EMD, DWT, RIIV, RIAV, RIFV, Fusion Freq. CapnoBase

Kozumplik et al., 2021 Rpc FFT*, RIIV, RIAV Freq., Time CapnoBase, BIDMC

Fikriastuti and

Muhaimin, 2021

Rpc FFT, RIIV, RIAV, RIFV, Fusion Freq. CapnoBase

Proto

saltis et al., 2021 R

c FFT*, RIIV, RIAV, RIFV, Fusion Fre

.Ca

noBase, Ad hoc

Haddad et al., 2021 Rpc RIIV, RIAV, RIFV, Fusion Time CapnoBase

Icazatti et al., 2021 Fl FIR Time BIDMC, CCSHS

Adami et al., 2021 Fl FFT*, EMD, DWT Fre

. BIDMC, MIT-BIH

Chen et al. 2021 R

c FFT*, RIIV, RIAV, RIFV, Fusion Fre

. Ad hoc

This wor

Fl, Rpc FFR*, Median, RIIV, Fusion Freq., Time BIDMC, CapnoBase

LEGEND: Rpc – Respiratory physiological characteristics; Fl – Only applying Filters

Freq. – Frequency domain; Time – Time domain

FFT – Fast Fourier Transform; FFT* – FFT used to extract the spectral frequency without capturing RR directly

RIIV, RIAV, RIFV – Respiratory-Induced Variation regarding Intensity, Amplitude and Frequency, respectively

Fusion – extract RR correlating RIIV, RIAV, RIFV; EMD – Empirical Mode Decomposition

EEMD, CEEMD, CEEMDAN, ICEEMDAN – Special methods based on EMD

DWT – Discrete Wavelet Transform; Median –

extracts RR from the baseline variation of the PPG signal median

that most of the works (Motin et al., 2018) (Motin et

al., 2019) (Motin et al., 2020) (Lazazzera and

Carrault, 2020) (Adami et al., 2021) extracts RR

using EMD and its variations (see column Method).

We identified the predominance of RIIV, RIAV,

and RIFV modulations in the analysis referring to the

extraction of RR based on respiratory physiological

characteristics, represented by Rpc in the Base

column.

Additionally, we divided the articles that extract

the RR based on the physiological characteristics of

breathing into three sets. These sets are composed of

works that assess the performance of RIIV, RIAV,

and RIIF: (i) only individually (Pollreisz and

TaheriNejad, 2020) (Lazazzera and Carrault, 2020)

(Kozumplik et al., 2021); (ii) individually, but

consider RR obtained with the highest quality index

(Khreis et al., 2018)(Birrenkott et al., 2018)(Khreis et

al., 2020); (iii) merging the partial RR values

obtained in each modulation to calculate the final RR

(Pimentel et al., 2017) (Orphanidou, 2017) (Yang,

2019) (Pollreisz and Nejad, 2020) (Lazazzera and

Carrault, 2020) (Fikriastuti, and Muhaimin, 2021)

(Protopsaltis et al., 2021) (Haddad et al., 2021) (Chen

et al., 2021).

The Domain column points out that most of the

works estimate RR by analyzing the signals in the

frequency domain, except for (Orphanidou, 2017)

(Khreis et al., 2018) (Motin et al., 2019) (Haddad et

al., 2021) (Icazatti et al., 2021) that employ the time

domain. Consequently, we explored the RR

extraction with methods that analyze both the time

and frequency domains, and we chose to employ the

Fast Fourier Transform (FFT) in all implemented

methods.

The Database column shows that most works use

CapnoBase, followed by the MIMIC and BIDMC

databases. This reason led us to choose the

CapnoBase and BIDMC databases in this work,

which are freely available and organized to obtain RR

from the PPG signal analysis. Additionally, we

decided to include the Synthetic Dataset from the

Methods to Estimate Respiratory Rate Using the Photoplethysmography Signal

447

Respiratory Rate Estimation project (Charlton, 2016)

to assess the ideal measurement time window, given

that data is in a noise-free environment.

The work proposed here developed three

algorithms entitled “FFT Method”, “Median

Method”, and “RIIV Method” to study the

performance when extracting the respiratory rate in

signals containing respiration-induced baseline

modulation. Besides, the work proposed here

implements the “Fusion Method”, which in its initial

stage evaluates RRs obtained with the three baseline

methods and merges these values when a given

criterion is accepted; otherwise, the measurement is

discarded.

4 METHODS FOR ESTIMATING

RESPIRATORY RATE

We implemented four methods to analyze the

effectiveness and efficiency of RR estimations from

the PPG signal. These methods enable estimating RR

(i) using an FFT; (ii) Median; (iii) exploring breath-

induced intensity variations, i.e., RIIV; and (iv)

integrating multiple RR estimates.

4.1 Fast Fourier Transform Method

The Fast Fourier Transform (FFT) quickly calculates

the discrete Fourier transform of a data sequence,

allowing us to obtain the signal frequency spectrum.

FFT is one of the most used methods to analyze

signals in the frequency domain because time is

mathematically eliminated during the transformation

process, resulting only in signal frequency

components (Tamura and Maeda, 2018).

This method uses the SciPyfft function of the

Python SciPy library to return the discrete Fourier

transform of a real or complex sequence (Bluestein,

1970). The extraction of RR from the PPG signal

follows the following steps:

 Converting the PPG signal from the time domain

to the frequency domain applying FFT;

 Transforming the frequency value to “rpm” by

multiplying the dominant peak value by 60;

 Identifying the dominant frequencies by

extracting the RR value within the acceptable

range for humans - between 4 rpm (0.06 Hz) and

60 rpm (1 Hz).

4.2 Median Method

The Median Filter allows extracting the baseline

present in the PPG signal (Awodeyi et al., 2013). The

Median Method explores the baseline variation by

calculating the PPG signal median.

This method uses the numpy.median function of

the Python Numpy library to return the median of the

vector elements. The RR extraction from the PPG

signal requires the following steps:

 Calculating the median with a sliding vector of the

signal sample rate size;

 Applying the FFT in the resulting vector and

selecting the dominant frequency peak within the

valid RR range;

 Multiplying the dominant frequencies by 60 to

convert to “rpm”.

4.3 RIIV Method

The RIIV Method explores baseline variations

obtained through linear interpolation between the

PPG signal peaks. We implemented this method

using the SciPy.find_peaks and SciPy.interpolate,

present in the Python SciPy library, which returns a

vector containing all the peaks of a signal and

function whose calling method interpolates to find the

value of new points (Bluestein, 1970).

RR can be acquired with an analysis in time or

frequency. Time analysis is achieved by counting

peaks or valleys in the signal. Frequency analysis

requires the FFT application; in this case, RR is the

dominant signal frequency generated by the FFT

within the RR human spectrum. The RR measured by

the respiration-induced variation methods is

performed via the following steps:

 Extraction of the induced baseline variation in the

PPG signal to get the respiratory signal in time;

 FFT Application on the respiratory signal and

selection of the dominant frequency peak within

the valid RR spectrum;

 Multiplying the dominant frequencies by 60 to

convert to “rpm”.

4.4 Fusion Method

Merging RR estimates complements the three

methods discussed earlier. This method ascends the

RR values obtained by the FFT, Median, and RIIV

methods, creating a three-position vector containing

the smallest (min), intermediate (itr), and largest

(max) values, and considers an error (ε) between

these values to determine the merge rule.

The determination of the error value is empirical,

and for this work, ε=0 rpm was chosen to perform an

in-depth analysis of the results.

The RR fusion estimate is constructed using the

HEALTHINF 2023 - 16th International Conference on Health Informatics

448

following conditions and considering their ordering:

 If the difference between the smallest and largest

vector value is less than or equal to the error, then

RR is the average of the three values. The reason

is to assume that if the three samples are close,

then they must be close to the true RR value;

 If the above conditions are not satisfied, RR is not

measured. As there is no consensus on the best

estimate, the method characterizes the sample as

a noisy signal, and the estimated value is

discarded. RR cannot be measured in an extreme

situation where this condition occurs in all

samples.

RR=

min + itr + max

min − max

≤ ε

not

evaluated

otherwise

5 EXPERIMENTAL

RESULTS – SYNTHETIC DATA

Synthetic Dataset is a database created in the

Respiratory Rate Estimation project that contains

synthetic ECG and PPG signals (Charlton, 2016).

These signals were developed to verify the

performance of methods to estimate RR in a scenario

without external noise. The dataset includes PPG

signals in the RR range between 4 and 60 rpm,

modulated in isolation concerning baseline,

amplitude, and frequency. It contains mathematically

equated PPG signals, not including external

interferences during its construction, enabling to get

accurate information on the influence of the window

size used for the RR measurement.

The initial length of the observation window was

empirically selected to be equal to 10s. Note that a

huge observation window requires a long time to

obtain a sample value; however, a tiny observation

window can lead to inaccuracy. The initial value was

considered the “smallest observation window” to

measure the RR. Subsequently, a value equivalent to

“smallest window” was added to find an observation

window size in which the absolute error between the

reference value and the estimated value remained

unchanged. This is considered the minimum precision

window; upper observation windows should not

change the RR results due to data subsampling.

All the methods analyzed in this section presented

stable values with an absolute error equal to zero for

observation windows greater than 50s. The

observation window of 10s was the only one that

presented an absolute error equal to 2 rpm, being

considered the highest error found. The window of

30s presented an absolute error equal to zero.

Some estimates obtained with 40s windows were

worse than the 30s estimates due to the signal being

synthetic, so the multiplicity of 40s generates a

precision error for the 500Hz samples. This finding

was essential to understand the influence of the

sampling rate in obtaining the best observation

window. Additionally, this finding indicates that it is

impossible to define an optimal observation window

value because it depends on the PPG signal.

6 EXPERIMENTAL

RESULTS – REAL DATA

The CapnoBase dataset (Karlen, 2010) contains ECG,

pulse oximetry with PPG, and CO

data from 42

cases, each described by an 8min recording. In

addition, CapnoBase provides the RR measured for

everyone, enabling us to evaluate the methods used.

BIDMC (Pimentel et al., 2018) is a dataset used to

assess the performance of RR methods. BIDMC

contains information from ECG, PPG pulse oximetry,

and respiratory signs from impedance

pneumography. All these signals were acquired from

53 intensive care patients, with recordings lasting

8min for each patient.

Figure 2 presents the configurations performed to

obtain the experimental data. The CapnoBase and

BIDMC databases underwent an equalization process

to obtain consistent RF reference values throughout

the sampling interval. The equalization generated the

banks called CapnoBase* and BIDMC*, which

contain sensory data from 42 and 53 people,

respectively; this data corresponds to 95 PPG signals,

each with 360s of sampling.

RR results

BIDMC*

Sliding observation

windows (30s, 60s)

632

60. 040

FFT

RIIV

RIAV

RIFV

EMD

Efficacy

Reference value

CapnoBase*

Equalization

CapnoBase

Equalization

BIDMC

Samples

Efficacy

Reference value

360. 240

Assessments

Fusion

60. 040

180. 120

60. 040

Figure 2: Settings performed to obtain experimental data.

These samples were evaluated by the FFT,

Median, RIIV, and Fusion methods in sliding

observation windows of 30s and 60s to assess

estimation errors with different window sizes.

Additionally, we chose to use sliding windows

with a base of 1s to smooth the passage from one

Methods to Estimate Respiratory Rate Using the Photoplethysmography Signal

449

sample to another – this approach was also adopted

by Fikriastuti and Muhaimin (2021). Using a sliding

window of 1s, within the range of 360s, implied 331

and 301 samples for the observation windows of 30s

and 60s, respectively. We chose to work with only six

minutes (360s) of the databases containing eight-

minute samples, removing the initial and final

minutes of each patient; this choice was made

empirically to remove the edges of the databases and

eliminate possible erroneous measurements.

This procedure results in 632 observation

windows for each of the 95 PPG signals;

consequently, 60,040 evaluations must be carried out

using three of the four proposed methods – except for

the Fusion Method, which takes the results of the

FFT, Median, and RIIV methods as input. The

execution of all experiments results in 360,240 RR

estimates, whose absolute error and corresponding

effectiveness are evaluated by comparing with the

reference values included in the CapnoBase and

BIDMC databases.

The Fusion Method evaluates the quality of the

signal when it integrates the three RR values

obtained—performing this method on CapnoBase

data removed 5722 and 4949 samples for observation

windows of 30s and 60s, respectively. Similarly,

BIDMC had 7493 and 6745 samples removed for

observation windows of 30s and 60s, respectively.

This removal is not directly related to the noise in the

database but instead to the discrepancy between the

FFT, Median, and RIIV methods, which did not allow

the Fusion Method to choose the best estimate to be

taken. Because of these removals, all the percentage

analyses presented in this section subtract the samples

classified as noisy from the total amount of samples

for the Fusion Method.

The two databases contain the reference RRs for

each PPG signal. However, in some cases, the

reference values to assess the CapnoBase RR were

only present in the range of 7s to 72s or smaller

intervals. In comparison, each patient has a total of

480s (8min) to store the PPG signal and its respective

respiratory reference signal (CO

). To obtain the RR

reference values over the entire 360s interval

explored in the database, we applied an FFT on the

signal.

To maintain similarity in the analysis of the two

databases, we applied the same procedure to the

Impedance signal present in the BIDMC database.

This process in both databases was called

Equalization. However, the Equalization approach

can fail when the analyzed window contains noise

that overlaps the respiratory signal, as illustrated in

Figure 3. Since the respiratory signal is not evaluated

Figure 3: Example of CapnoBase data containing intense

noise that overlaps with (a) PPG and (b) CO

signal, and (c)

RR reference values.

correctly, the RR estimates will likely suffer

deviations that can alter the efficacy analysis of the

evaluated methods.

We decided to evaluate two window observation

sizes (30s and 60s), as the variability contained in the

databases can generate artifacts that can be minimized

in larger windows. Finally, we chose to work with

only 6 minutes of the databases containing eight-

minute samples, discarding the initial and final

minutes of each patient. Figure shows the percentage

of samples with 100% accuracy in the RR obtained

with each evaluated method, considering the data

provided by CapnoBase and BIDMC for both

observation windows. Regarding only the CapnoBase

dataset, the result analysis shows an average variation

of practically 5% error with the observation window

reduction. This variation indicates that the proposed

methods improve the RR estimates as the observation

window increases. Additionally, the Fusion Method

obtains the highest efficacy, improving the quality of

the estimates obtained with the other methods.

Figure 4: Percentage of samples reaching 100% efficacy

using CapnoBase and BIDMC datasets and sliding

observation windows of 30s and 60s.

The results of BIDMC indicate many similarities

in the data contained in the databases. Clear examples

62.28%

60.53%

54.51%

78.98%

60.95%

58.72%

54.22%

78.86%

70.21%

77.05%

65.28%

90.24%

75.03%

78.18%

66.82%

92.93%

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

100%

FFT MEDIAN RIIV FUSION

BIDMC (30s) BIDMC (60s)

CAPNOBASE (30s) CAPNOBASE (60s)

HEALTHINF 2023 - 16th International Conference on Health Informatics

450

are the Fusion Method, which had the best results for

the BIDMC and CapnoBase databases, and the RIIV

method, which had the lowest estimates in both

databases. Additionally, the 60s observation window

performed poorly compared to the results obtained

with 30s windows.

Figure presents the same experiment as Figure

but highlights the number of RR samples that had

errors of 3 rpm or more. Once again, it is possible to

observe that the Fusion Method achieves the best

results for both databases and observation windows,

with the worst results obtained with the RIIV Method.

Although RR estimates with errors of 3 rpm or more

may be unacceptable, the number of samples with

these values is proportionally low when using Fusion

Method. Additionally, samples with high errors are

interlaced with samples of greater efficacy, enabling

filtering software to remove sudden variations

between samples and improving the average accuracy

of the RR estimate.

Figure 5: Percentage of samples with errors upper to 3 rpm

of RR using CapnoBase and BIDMC datasets and sliding

observation windows of 30s and 60s.

Additionally, Figure and Figure display that all

methods are more effective when working with the

CapnoBase database, and Fusion and RIIV methods

have the best and worst performances among the

methods, respectively. However, the same efficacy

was not achieved with the BIDMC database. A

possible reason for this achievement is acquired when

visually analyzing the PPG signals; i.e., a higher

number of samples with artifacts was identified in the

BIDMC database, which is the main reason for the

methods having lower performance compared to the

evaluations made with CapnoBase.

7 CONCLUSION

This work evaluated the effectiveness of four

methods for inferring the respiratory rate of PPG

signals available in synthetic databases. The Fusion

Method showed the highest efficiency among the

evaluated methods, mainly because the merger

achieved, on average, the best results of the correlated

methods. The proposed methods can be easily

implemented within firmware on a wearable device,

making it possible to evaluate the methods explored

in this work in real scenarios and in real-time. As a

future work, we intend to compare the methods

proposed here with other works described in the

literature to assess the relative effectiveness of the

approaches.

ACKNOWLEDGMENT

This study was financed in part by the Coordination

for the Improvement of Higher Education Personnel -

Brazil (CAPES) - Finance Code 001, National

Council for Scientific and Technological

Development (CNPq) and Financier of Studies and

Projects (FINEP).

REFERENCES

Olivadoti, G. (2022). How Advances in Sensor and Digital

Technology Yield Better Patient Care, Thought

Leadership Article, Analog Devices, pp. 1-2.

Charlton, P., Birrenkott, D., Bonnici, T., Pimentel, M.,

Johnson, A., Alastruey, J., Tarassenko, L., Watkinson,

P., Beale, R., Clifton, D. (2018). Breathing Rate

Estimation from the Electrocardiogram and

Photoplethysmogram: A Review, IEEE Reviews in

Biomedical Engineering, v. 11, pp. 2-20.

Karlen, W., Turner, M., Cooke, E., Dumont, G.,

Ansermino, J., (2010). CapnoBase: Signal database and

tools to collect, share and annotate respiratory signals,

Proceedings of the Annual Meeting of the Society for

Technology in Anesthesia (STA), pp. 1-27.

Pimentel, M., Johnson, A., Charlton, P., Clifton, D. (2018).

BIDMC PPG and Respiration Dataset,

https://physionet.org/content/bidmc/1.0.0/.

Kamal, A., Harness, J., Irving, G., Mearns, A. (1989). Skin

photoplethysmography—a review, Computer methods

and programs in biomedicine, v. 28, n. 4, pp. 257-269.

Tamura, T., Maeda, Y., Sekine, M., Yoshida, M. (2014).

Wearable photoplethysmographic sensors—past and

present, Electronics, v. 3, n. 2, pp. 282-302.

Mejıa-Mejıa, E., Allen, J., Budidha, K., El-Hajj, C.,

Kyriacou, P., Charlton, P. (2021).

Photoplethysmography Signal Processing and

Synthesis”, Photoplethysmography - Technology,

Signal Analysis and Applications, Academic Press,

chap. 4, pp. 69-146.

30.43%

32.45%

38.23%

16.83%

24.57%

27.56%

32.97%

11.82%

25.36%

18.64%

27.02%

7.74%

20.72%

16.59%

24.08%

5.55%

10%

15%

20%

25%

30%

35%

40%

45%

FFT MEDIAN RIIV FUSION

BIDMC (30s) BIDMC (60s)

CAPNOBASE (30s) CAPNOBASE (60s)

Methods to Estimate Respiratory Rate Using the Photoplethysmography Signal

451

Tamura, T. and Maeda, Y. (2018) Pulse and Flow –

Photoplethysmogram, Seamless Healthcare

Monitoring - Advancements in Wearable, Attachable,

and Invisible Devices, Springer, Part III, pp. 159-192.

Dehkordi, P., Garde, A., Molavi, B., Ansermino, J.,

Dumon, G. (2018). Extracting Instantaneous

Respiratory Rate from Multiple Photoplethysmogram

Respiratory-Induced Variations, Frontiers in

Physiology, v. 9, pp. 948.1-948.10.

Charlton, P., Bonnici, T., Tarassenko, L., Clifton, D., Beale,

R., Watkinson, P. (2016). An Assessment of

Algorithms to Estimate Respiratory Rate from The

Electrocardiogram and Photoplethysmogram,

Physiological Measurement, v. 37, n. 4, pp. 610-626.

Pimentel, M., Johnson, A., Charlton, P., Birrenkott, D.,

Watkinson, P., Tarassenko, L. Clifton, D. (2017).

Toward a Robust Estimation of Respiratory Rate from

Pulse Oximeters, IEEE Transactions on Biomedical

Engineering, v. 64, n. 8, pp. 1914-1923.

Orphanidou, C. (2017). Derivation of respiration rate from

ambulatory ECG and PPG using Ensemble Empirical

Mode Decomposition: Comparison and fusion,

Computers in Biology and Medicine, v. 81, pp. 45-54.

Motin, M., Karmakar, C. Palaniswami, M. (2018)

Ensemble Empirical Mode Decomposition with

Principal Component Analysis: A Novel Approach for

Extracting Respiratory Rate and Heart Rate from

Photoplethysmographic Signal, IEEE Journal of

Biomedical and Health Informatics, v. 22, n. 3, pp. 766-

774.

Khreis, S., Ge, D., Carrault, G. (2018). Estimation of

Breathing Rate from the Photoplethysmography Using

Respiratory Quality Indexes, Proceedings of the

Computing in Cardiology Conference (CinC), pp. 1-4.

Birrenkott, D., Pimentel, M., Watkinson, P., Clifton, D.

(2018). A Robust Fusion Model for Estimating

Respiratory Rate from Photoplethysmography and

Electrocardiography, IEEE Transactions on

Biomedical Engineering, v. 65, n. 9, pp. 2033-2041.

Motin, M., Karmakar, C., Palaniswami. M. (2019). Selection

of Empirical Mode Decomposition Techniques for

Extracting Breathing Rate From PPG, IEEE Signal

Processing Letters, v. 26, n. 4, pp. 592-596.

Yang, H., Li, M., He, D., Che, X., Qin, X. (2019).

Respiratory Rate Estimation from the

Photoplethysmogram Combining Multiple

Respiratory-induced Variations Based on SQI,

Proceedings of the IEEE International Conference on

Mechatronics and Automation (ICMA), pp. 382-387.

Pollreisz D. and Nejad, N. (2020). Reliable Respiratory

Rate Extraction using PPG, Proceedings of the IEEE

Latin American Symposium on Circuits & Systems

(LASCAS), pp. 1-4.

Motin, M., Karmakar, C., Kumar, D., Palaniswami, M.,

(2020). PPG Derived Respiratory Rate Estimation in

Daily living Conditions, Proceedings of the Annual

International Conference of the IEEE Engineering in

Medicine & Biology Society (EMBC), pp. 2736-2739.

Pollreisz, D. and TaheriNejad, N. (2020). Efficient

Respiratory Rate Extraction on a Smartwatch,

Proceedings of the Annual International Conference of

the IEEE Engineering in Medicine & Biology Society

(EMBC), pp. 5988-5991.

Khreis, S., Ge, D., Rahman, H., Carrault, G. (2020).

Breathing Rate Estimation Using Kalman Smoother

with Electrocardiogram and Photoplethysmogram,

IEEE Transactions on Biomedical Engineering, v. 67,

n. 3, pp. 893-904.

Lazazzera, R. and Carrault, G. (2020). Breathing Rate

Estimation Methods from PPG Signals, on CapnoBase

Database, Proceedings of the Computing in Cardiology

(CinC), v. 47, pp. 1-4.

Kozumplik, J., Smital, L., Nemcova, A., Ronzhina, M.,

Smisek, R., Marsanova, L., Kralik, M., Vitek, M.

(2021). Respiratory Rate Estimation Using the

Photoplethysmogram: Towards the Implementation in

Wearables, Proceedings of the Computing in

Cardiology (CinC), v. 48, pp. 1-4.

Fikriastuti, N. and Muhaimin, H. (2021). Respiratory Rate

Estimations using Three Respiratory-Induced

Variations on Photoplethysmogram, Proceedings of the

International Conference on Electrical Engineering

and Informatics (ICEEI), pp. 1-6.

Protopsaltis, G., Krizea, M., Gialelis, J. (2021). Continuous

Measurement of Respiratory Rate via Single-

Wavelength Reflective Photoplethysmography,

Proceedings of the Mediterranean Conference on

Embedded Computing (MECO), pp. 1-5.

Haddad, S., Boukhayma, A., Caizzone, A. (2021). PPG-

Based Respiratory Rate Monitoring Using Hybrid

Vote-Aggregate Fusion Technique, Proceedings of the

Annual International Conference of the IEEE

Engineering in Medicine & Biology Society (EMBC),

pp. 1605-1608.

Icazatti, F., Dell’Aquila, C., Leber, E. (2021). Design and

validation of a respiratory rate estimation algorithm

based on photoplethysmography (PPG) signal,

Proceedings of the Workshop on Information

Processing and Control (RPIC), pp. 1-5.

Adami, A., Boostani, R., Marzbanrad, F., Charlton, P.

(2021). A New Framework to Estimate Breathing Rate

from Electrocardiogram, Photoplethysmogram, and

Blood Pressure Signals, IEEE Access, v. 9, pp. 45832-

45844.

Chen, M., Zhu, Q., Wu, M. and Wang, Q. (2021).

Modulation Model of the Photoplethysmography

Signal for Vital Sign Extraction, IEEE Journal of

Biomedical and Health Informatics, v. 25, n. 4, pp. 969-

977.

Bluestein, L. (1970). A linear filtering approach to the

computation of discrete Fourier transform. IEEE

Transactions on Audio and Electroacoustics, v. 18, n.

4, pp. 451-455.

Awodeyi, A., Alty, S. and Ghavami, M. (2013). Median

Filter Approach for Removal of Baseline Wander in

Photoplethysmography Signals, European Modelling

Symposium, pp. 261-264.

HEALTHINF 2023 - 16th International Conference on Health Informatics

452